WO2022210361A1

WO2022210361A1 - Analyzing device and analyzing method

Info

Publication number: WO2022210361A1
Application number: PCT/JP2022/014421
Authority: WO
Inventors: 哲也金田
Original assignee: 株式会社D’isum
Priority date: 2021-04-02
Filing date: 2022-03-25
Publication date: 2022-10-06

Abstract

The purpose of the present invention is to make it possible to check whether or not there is a possibility of abnormality from numerical data including frequency components of machine or human motion.　The present invention provides an analyzing device. The analyzing device stores a set of reference vector data obtained by converting data concerning a subject involving temporal motion into multidimensional vector data including frequency as a dimension, calculates a reference vector for a multidimensional space defined by the set of reference vector data, converts, upon acquiring new data concerning the subject, the new data into new multidimensional vector data including frequency as a dimension, and determines, by using the reference vector, whether or not the position of the new multidimensional vector data in the multidimensional space falls within the region of the multidimensional space defined by the set of reference vector data.

Description

Analysis device and analysis method

The present disclosure relates to an analysis apparatus and an analysis method for converting the movement of an object into numerical data, which is its frequency component, and analyzing it.

A method has been proposed for determining abnormalities in an object that moves with time, such as a mobile object (see Patent Document 1, for example). In Patent Literature 1, frequency component data is categorized into normal and abnormal by clustering processing. For category classification, it is necessary to learn in advance the relationship between frequency components and categories.

In the case of numerical data with frequency components, it is difficult to extract common characteristic frequency components. For example, even machines of the same type have individuality, and the combination of characteristic frequency components differs from machine to machine. Therefore, it is very difficult to learn in advance the relationship between frequency components and categories.

JP 2012-58171 A

The purpose of the present disclosure is to make it possible to determine the possibility of an abnormality from numerical data having frequency components of movements of machines and people.

The analysis device and analysis method of the present disclosure are
Holds a set of reference vector data obtained by converting target data with temporal movement into multidimensional vector data including frequency as a dimension,
calculating a reference vector in a multidimensional space defined by the set of reference vector data;
When the new data of the target is obtained, the new data is converted into new multidimensional vector data including frequency as a dimension;
Using the reference vector, it is determined whether the position of the new multidimensional vector data in multidimensional space is within the region of the multidimensional space defined by the set of reference vector data.

The program of the present disclosure is a program for realizing a computer as each functional unit provided in the apparatus according to the present disclosure, and is a program for causing the computer to execute each step included in the method executed by the apparatus according to the present disclosure. .

According to the present disclosure, it is possible to determine whether or not there is a possibility of abnormality even in numerical data having frequency components.

1 shows a system configuration example of the present disclosure; An example of the operation of the analysis device is shown. It is an example of a cluster in a multidimensional space and vector data that has begun to deviate from the cluster. An example of an algorithm for determining deviation from a cluster is shown. An example of reference vector data set {R _i }, reference vector G and new multidimensional vector data X is shown. FIG. 2 is an example of a two-dimensional representation of the distribution of multidimensional vector data of bearing vibration, that is, a reference map. FIG. 4 is a flow chart showing an example of a data visualization method; A plot example of new multidimensional vector data X on the reference map is shown. 3 shows an example of distribution of multidimensional vector data of SAS. An example of multidimensional vector data of a plurality of motors provided in one machine is shown. An example of a state map is shown.

Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. Note that the present disclosure is not limited to the embodiments shown below. These implementation examples are merely illustrative, and the present disclosure can be implemented in various modified and improved forms based on the knowledge of those skilled in the art. In addition, in this specification and the drawings, constituent elements having the same reference numerals are the same as each other.

(First embodiment)
FIG. 1 shows an example of the system configuration of the present disclosure. The analysis device 10 of the present disclosure includes a storage unit 11 that stores data and a signal processing unit 12 . The analysis device 10 may include a communication section 13 and a display section 14 . The analysis device 10 of the present disclosure can also be realized by a computer and a program, and the program can be recorded on a recording medium or provided through a network.

(target data)
The present disclosure deals with data of temporally moving objects. For example, machine vibration and sound data detected by the sensor 20 can be exemplified. Since motion can be represented by frequency, data of an object with temporal motion can be converted into multi-dimensional vector data whose dimension is frequency. The data targeted in the present disclosure is not limited to vibrations and sounds generated by machines, and can be applied to arbitrary data that can move, such as people and automobiles.

The analysis device 10 acquires data of an object that moves temporally and stores it in the storage unit 11 . The target data may be acquired from the communication network 100 via the communication unit 13, but the present disclosure is not limited to this.

FIG. 2 shows an example of the operation of the analysis device 10. As shown in FIG.
The analysis device 10 generates a set {R _i } of reference vector data from target data (S01),
Calculate the reference vector G of {R _i } on the multidimensional space,
When new target data is acquired (S03),
generating new multidimensional vector data X from new target data (S04);
Using the reference vector G, the relationship between the reference vector data set {R _i } and the new multidimensional vector data X is determined (S05).

(Generation of reference vector data S01)
The storage unit 11 stores data of an object that temporally moves. The signal processing unit 12 divides the data into multiple segments. The division is by time relative to the data. As a result, a plurality of segments are generated by temporally dividing the motion. The signal processing unit 12 generates multidimensional vector data having frequency components as dimensions for each segment (S01). As a result, a reference vector data set {R _i } is generated and stored in the storage unit 11 . where i is the identifier of each segment.

When an object that moves temporally is normal, the reference vector data forms one or more clusters in the multidimensional space, as shown in the distribution of normal data in FIG. In each cluster, the data contained in the cluster generally exhibits a recursive movement on each dimension during normal operation, and therefore has a mountain-shaped distribution such as Gaussian distribution or binomial distribution.

(Calculation S02 of reference vector G)
The signal processing unit 12 calculates a reference vector G in a multidimensional space defined by the reference vector data set {R _i } (S02). The reference vector G is, for example, a centroid vector of a multidimensional vector composed of a set {R _i } of reference vector data.

(Acquisition of new data and conversion to multidimensional vector data S03, S04)
When new target data is acquired (S03), frequency components are extracted and converted into multidimensional vector data having the frequency components as dimensions. As a result, new multidimensional vector data X is generated (S04).

(Relationship determination S05)
When the determination target is normal, the new multidimensional vector data X is positioned within the cluster formed by the reference vector data set {R _i }. When some kind of abnormality occurs in the determination target, the new multidimensional vector data X begins to deviate from the normal data distribution area, like the abnormal data indicated by ▴ in FIG. However, in FIG. 3, the multidimensional data is represented on two dimensions using the dimensionality reduction technique. Therefore, the signal processing unit 12 determines the relationship between the reference vector data set {R _i } and the new multidimensional vector data X. FIG. For example, the signal processing unit 12 determines whether the position of the new multidimensional vector data X in the multidimensional space is within the region of the multidimensional space defined by the reference vector data set {R _i }.

If the position of the new multidimensional vector data X in the multidimensional space is outside the region defined by the reference vector data set {R _i }, the signal processing unit 12 determines the target corresponding to the new multidimensional vector data X. Determine that the state is different from the state of interest defined by the reference vector data set {R _i }.

FIG. 4 shows an example of an algorithm executed by the signal processing section 12 in step S05.
First, we deal with the case where there is one cluster in the reference vector data set {R _i }.
The signal processing unit 12 calculates the reference vector G of the reference vector data set {R _i } (S11).
Next, the signal processing unit 12 uses the reference vector G to determine whether the position of the new multidimensional vector data X in the multidimensional space is within the multidimensional space region of the reference vector data set {R _i }. It is determined whether or not (S12).

FIG. 5 shows an example of the reference vector data set {R _i }, the reference vector G and the new multidimensional vector data X. In FIG. Whether or not X is within the region of {R _i } is determined by multiplying the new multidimensional vector data X by the vector (R _i −G) connecting the multidimensional vector R _i included in the reference vector data and the reference vector G. The component in the same direction as the vector (XG) connecting the dimension vector and the reference vector G can be compared with the length |XG| of the vector (XG), and the magnitude relation can be used for determination. .

For example, the signal processing unit 12 determines that the new multidimensional vector data X is within the cluster if (XG, R _i -G)≧|XG| ² for all i included in the cluster. can do. where (x, y) is the inner product of x, y. On the other hand, the signal processing unit 12 determines that the new multidimensional vector data X is outside the cluster if (XG, R _i -G)≦|XG| ² for all i included in the cluster. (S16). where (x, y) is the inner product of x, y.

The signal processing unit 12 may calculate a divergence index ρ indicating how much the new multidimensional vector data X diverges from the reference vector data set {R _i }. Assuming that R _i at which (XG, R _i -G) is maximum is R _max , the divergence index ρ can be calculated using, for example, the following equation.
(Number 1)
ρ=|X−G| ² /(X−G, R _max −G)−1 (1)

When using this formula, the signal processing unit 12 can determine that it is outside the cluster if ρ>0, and it can determine that it is inside the cluster if ρ<0. The larger ρ is, the more it deviates from the cluster. Therefore, the present disclosure can quantify the divergence index ρ. In actual operation, a determination threshold _ρ _th is set in order to _allow for a certain amount of error. It would be common to assume that

Next, consider a case where there are a plurality of clusters in the reference vector data set {R _i }.
In that case, the simplest method is a method of "regarding a plurality of clusters as one cluster". When an anomaly occurs, new data generally begin to deviate from any cluster. Therefore, even if a plurality of clusters are regarded as one cluster, at least part of the data can be judged to be abnormal by the above judgment method.

On the other hand, it is possible to improve the judgment accuracy by judging the abnormality of new data for each cluster. In that case, the above method can be applied to each cluster by calculating the reference vector G for each cluster in step S11. This method is effective when a plurality of clusters are clearly separated, but cannot be said to be effective when the clusters overlap. For reference, it would be realistic to focus on one cluster and apply this method, even if it does not apply to all previous clusters.
When the target is an actual machine or the like, it is considered realistic to apply the simple method described above.

Thus, the present disclosure determines the relationship between the set of reference vector data {R _i } and the new multidimensional vector data X using a geometric algorithm using the coordinates of the data in the multidimensional space. do. As a result, the degree of freedom in designing the customer's operation system is greatly improved, and the cost can be greatly reduced.

(Second embodiment)
In the first embodiment, the reference vector G is the centroid of the multidimensional vector data set {R _i }. Here, a case is shown in which the reference vector G is set in the outer region of the multidimensional vector data set {R _i }. Specifically, the origin O of the multidimensional vector data forming the reference vector data is used as the reference vector.

In the case of objects with monotonous motion, such as bearings, the deterioration is almost always caused by an increase in the amplitude of the motion. Although the change in amplitude is slight at the beginning of deterioration, the amplitude increases as deterioration progresses. At the same time, the shape of the spectrum also changes. In order to detect such movements, it is effective to use the origin, ie, the point where all frequency components are zero, as the reference vector.

An example of the distribution of bearing vibration data is shown in FIG. In the case of bearings, the amplitude of vibration is constant under normal conditions, so data is distributed on the surface of a multidimensional sphere in multidimensional space. The spread of the distribution represents the variation of the frequency spectrum. FIG. 6 shows the distribution of this data two-dimensionally.

By using normal data as the reference vector data, it is possible to determine whether the new multidimensional vector data X is abnormal and to evaluate the deviation from the normal data. Specifically, the divergence index can be defined as follows in the same manner as in Equation (1).
(Number 2)
ρ=|X| ² /(X, R _max )−1 (2)
Here, in the present embodiment, let {R _i } be a set of normal reference vector data, and let R _max be the R _i with the maximum inner product (X, R _i ).

If there is a sign of abnormality in the new multidimensional vector data X, a display to the effect that there is a possibility of abnormality may be displayed on the display unit 14 . It should be noted that the stage at which the warning is issued is arbitrary, and the threshold for issuing the warning may be settable. The divergence index ρ quantifies the degree of actual deterioration. Therefore, an alarm may be issued according to the divergence index ρ.

In this embodiment, it is possible to determine normality/abnormality using new data with a simple algorithm and in a short time. Both the mapping of the additional data and the calculation of the divergence index ρ can be easily calculated with a normal computer. It can also be easily integrated into existing machine operating systems.

(Third embodiment)
In the field of maintenance of actual machinery and equipment, the effect of visualizing the deterioration of equipment is great. However, since the data in the multidimensional space cannot be visualized as it is, it must be displayed in two dimensions, and an example of this method will be described here.

The signal processing unit 12 plots the reference vector data set {R _i } on a two-dimensional plane to create a reference map. Furthermore, when new multidimensional vector data X is obtained, it is plotted on the reference map. The overall algorithm is explained in FIG. Steps S01 to S05 shown in FIG. 7 are as described in the first embodiment.

When the signal processing unit 12 generates the reference vector data set {R _i } (S01), the number of dimensions of the multidimensional reference vector data is reduced to two, and the reference vector data set {R _i } is reduced to two dimensions. A reference map plotted on a plane is created (S21).
Then, when the new multidimensional vector data X is generated (S04), the signal processing unit 12 calculates the plot position of the new multidimensional vector data X on the reference map (S22), and generates new multidimensional vector data X on the reference map. A two-dimensional map on which the dimensional vector data X is plotted is displayed on the display unit 14 (S23).

As a method of plotting the reference vector data set {R _i } on a two-dimensional plane, toorPIA or t-SNE (T-distRibated Stochastic Neighbor Embedding), machine learning, or any other dimensionality reduction algorithm can be used.

The method of plotting the two-dimensional vector data χ corresponding to the new multidimensional vector data X on the reference map is arbitrary, and various methods can be adopted.

One method is to maintain the geometric relationship between the reference vectors G, R _max and X in the multidimensional space on the two-dimensional map, and to point g and R _max on the reference map corresponding to G This is a method of determining the point of the two-dimensional vector data χ on the X reference map from the position of the corresponding point r _max on the reference map.

As another method, R _near that is angularly closest to the new multidimensional vector data X in the multidimensional space is extracted from the reference vector data set {R _i }. Specifically, if the point on the reference map corresponding to R _near is r _near , the position of the two-dimensional vector data χ on the reference map corresponding to X as shown in FIG. 8 is expressed below.

Here, r _near is a point corresponding to R _near on the reference map.

There are many other possible methods for plotting the two-dimensional vector data χ on the reference map. In any case, since it is originally impossible to accurately reproduce the position of the vector on the multidimensional space on the two-dimensional reference map, any method is an approximation method.

By displaying new multi-dimensional vector data X using a two-dimensional reference map, it becomes possible to grasp changes in the state of the object to be judged in real time, and to visually grasp the state of deterioration of the object to be judged in real time. can be

(Fourth embodiment)
This technology can also be applied to detecting abnormalities in biological information and grasping conditions. Sleep apnea syndrome (SAS: Sleep Apnea Syndrome) affects hundreds of millions of people around the world, and is said to increase the risk of accidents such as traffic accidents and the risk of death in patients with underlying diseases by 5 to 7 times. ing. At present, a major obstacle to the treatment of SAS is that simple examinations, which are the first step in determining the suitability of treatment, are troublesome and time consuming. This technology is effective in realizing a simple method that replaces this simple inspection.

Specifically, the analysis apparatus 10 of the present embodiment acquires breath sounds recorded by a smartphone or the like during sleep, and divides the breath sounds into segments of about one minute, which are predetermined time widths. , segment by segment into frequency components and compute spectral data. Spectral data is multidimensional vector data, and segment data is distributed in multidimensional space.

From respiratory sound data during sleep of patients and healthy people, it is possible to obtain a multidimensional spatial distribution of standard respiratory sound segments. Here, the distribution range of the segment data of the healthy person is assumed to be the distribution range of the normal respiratory data. When a new subject's breath sound data is obtained, it is arranged at one point in this multidimensional space by converting it into vector data for each segment.

For example, by using the center of gravity of the normal respiration data as a reference vector, it is possible to determine whether the subject data is within the distribution range of the normal respiration data (normal segment) or outside the distribution range (abnormal segment). An indication of the subject's degree of SAS is obtained based on what percentage of the subject's total segment data in a given period of time is normal segments. Here, in actual operation, an area that is wider (or narrower) than the normal area to some extent may be used to determine the abnormal segment in order to determine that the subject's segment is abnormal.

Fig. 9 shows an example distribution of segment data for patients with various types of SAS and those without symptoms of SAS. Here, the center of gravity of the normal respiration data is used as a reference vector, and the multidimensional space data is displayed two-dimensionally by dimensionality reduction. Segment abnormal breathing patterns include "obstructive type", "central type", "hypopnea type", and "mixed type", and in general, each type of segment is distributed in a different area in multidimensional space. can be done. A subject's SAS type can also be determined based on which region the subject's segments are distributed.

(Fifth embodiment)
In a machine (for example, a robot) that incorporates a plurality of motors, it is possible to detect signs of an abnormality in the machine by monitoring the load on each motor. Specifically, the analysis device 10 of the present embodiment acquires load information such as current, voltage and torque of each motor provided in the machine, vibration and sound information, and the like, and converts them into multidimensional vector data. As a result, in this embodiment, the state of the mechanical equipment is grasped, and when the change in the state satisfies a certain condition, it is determined to be "abnormal".

Fig. 10 shows an example of multidimensional vector data of each motor provided in one machine. The figure shows an example of plotting on a two-dimensional plane according to the degree of similarity of each piece of multidimensional vector data. When the machine is equipped with six motors, the multidimensional vector data of each motor is distributed in defined areas M11, M21, M31, M41, M51, M61 in the multidimensional space. A set of multi-dimensional vector data of each motor thus obtained becomes reference vector data of each motor.

An example of the operation of the analysis device 10 of this embodiment is shown.
Step 1) Create reference vector data for a plurality of motors for each machine (Fig. 10).
Step 2) When new target data is acquired, the "degree of divergence" of the new data from the reference vector data is calculated for each machine and each motor.
Step 3) Six-dimensional vector data whose dimension is the "degree of divergence" for each motor is created for each machine. In this embodiment, this 6-dimensional vector data is referred to as state vector data of the machine.
Step 4) If there are multiple machines, use each machine's set of state vector data to create a plane map. In this embodiment, this plane map is called a state map. A value of each dimension of the state vector data represents an abnormality of each motor. For this reason, when an abnormality occurs in one of the motors, the state vector data of the machine including that motor moves away from the center, but is plotted in different directions when viewed from the center depending on which motor has deteriorated further. (See Fig. 11)
Step 5) Using the state vector data, it is visualized which of the 6 motors is worse than the others.

Furthermore, as shown in FIG. 11, by defining a normal region R1, a quasi-normal region R2, a caution region R3, an abnormal region R4, etc. on this state map, the state of the entire mechanical equipment can be visually grasped. can be done. Of course, by clicking one piece of machinery on the state map, it is possible to further display the state map of each motor of that piece of machinery.

This disclosure can be applied to the information and communications industry.

10: Analysis device 11: Storage unit 12: Signal processing unit 13: Communication unit 14: Display unit 20: Sensor 100: Communication network

Claims

Holds a set of reference vector data obtained by converting target data with temporal movement into multidimensional vector data including frequency as a dimension,
calculating a reference vector in a multidimensional space defined by the set of reference vector data;
When the new data of the target is obtained, the new data is converted into new multidimensional vector data including frequency as a dimension;
Using the reference vector, determining whether the position of the new multidimensional vector data in the multidimensional space is within a region of the multidimensional space defined by the set of the reference vector data;
analysis equipment.
When the position of the new multidimensional vector data in the multidimensional space is determined to be outside the region, the state of the object corresponding to the new multidimensional vector data corresponds to the set of reference vector data. Determine that it is different from the state of the target that
The analysis device according to claim 1.
calculating a divergence index between the set of reference vector data and the new multidimensional vector data using the reference vector, the set of reference vector data, and the new multidimensional vector data;
The analysis device according to claim 1.
The magnitude of the vector connecting the multidimensional vector of the reference vector data and the reference vector in the same direction as the vector connecting the multidimensional vector of the new multidimensional vector data and the reference vector is determined by the new multidimensional vector. calculating the divergence index by comparing the magnitude of the vector connecting the vector data and the reference vector;
The analysis device according to claim 3.
wherein the reference vector is the centroid of the set of reference vector data in multidimensional space;
The analysis device according to claim 1.
The target data is breath sound data,
the condition of the subject includes no symptoms of sleep apnea;
calculating the reference vector for no symptoms of sleep apnea;
When new breath sound data is acquired, it is converted into new multidimensional vector data,
using the reference vector to determine whether the new multidimensional vector data in multidimensional space is either with or without symptoms of sleep apnea;
The analysis device according to claim 1.
The target data is data obtained by dividing respiratory sound data into segments of a predetermined time width and then frequency-converting each segment,
The subject segment status includes obstructive, central or hypopnea sleep apnea, and no symptoms of sleep apnea;
calculating the reference vector for each set of segments of obstructive, central or hypopnea sleep apnea, and no symptoms of sleep apnea;
When new breath sound data is acquired, it is divided into segments of a predetermined time width, and then frequency-transformed for each segment to generate multidimensional segment vector data,
Using each of the reference vectors, the new multidimensional segment vector data in multidimensional space are obstructive, central, or hypopnea sleep apnea, and no symptoms of sleep apnea. determine whether either
The analysis device according to claim 1.
The target is a composite target including multiple moving targets,
calculating the divergence index for each of the targets;
generating state vector data whose dimension is the divergence index for each of the composite targets;
4. The analysis according to claim 3, wherein the set of multiple complex objects is displayed on a state map by plotting on a two-dimensional plane according to the mutual similarity of the state vector data of the complex objects. Device.
Holds a set of reference vector data obtained by converting target data with temporal movement into multidimensional vector data including frequency as a dimension,
calculating a reference vector in a multidimensional space defined by the set of reference vector data;
When the new data of the target is obtained, the new data is converted into new multidimensional vector data including frequency as a dimension;
Using the reference vector, determining whether the position of the new multidimensional vector data in the multidimensional space is within a region of the multidimensional space defined by the set of the reference vector data;
analysis method.